Dear Allstat Community,In a recent post on my blog, I demonstrated how logarithmic transformation can linearize a target-predictor relationship.I would much appreciate your help with 2 sets of difficulties.1) If you plot the residuals against the fitted values of ln(DDT), you will find that the variance decreases. I am encountering 2 problems:1a) I can't find a good transformation to remove this trend in the variance.1b) I have already transformed the model once. I fear that transforming it again would make it difficult to make inferences on the quantities and obtain an easily understandable interpretation of the quantities.2) At the end of the blog post, I used the estimated decay rate to calculate the half-life. I would like to calculate a confidence interval for the half-life. Even if I (incorrectly) set aside the non-constant variance that ruins my estimation of the standard errors, I don't know how to calculate a confidence interval for the half-life, since the variance calculation would start off asVar[ln(0.5)/lambda] = Var[half-life],where lambda is the random variable in this case; its standard error is given in my regression output for beta1. Half-life is not a linear combination of lambda, so the variance cannot be nicely calculated with the usual rules of variance. Thus, I don't know how to calculate the variance of half-life and, hence, the confidence interval for half-life. (I fully understand that the standard error for beta1/lambda is not good, so I welcome any suggestion for overcoming both of these problems simultaneously.)I would much appreciate your insights on how I can overcome these problems.Thank you for your time.Eric________________________Eric CaiThe Chemical StatisticianTwitter: @chemstatericM.Sc.StatisticsUniversity of TorontoB.Sc. with DistinctionChemistry and MathematicsSimon Fraser University
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.